Agent-tooling installers: managed CLAUDE.md blocks vs clobbered INSTRUCTIONS files

When an agent-integration CLI installs itself into a coding-agent config, it usually treats two files very differently, and conflating them wastes debugging time. The generated instructions file (e.g. ~/.config/<tool>/INSTRUCTIONS.md) is typically an unconditional writeFileSync on every install — no merge, no backup — so re-running install always refreshes it but silently destroys local edits. The agent memory file (CLAUDE.md) instead gets a delimiter-managed block: the installer cuts everything between <!-- tool:start --> and <!-- tool:end --> and re-appends fresh, so hand-written prose elsewhere in that file is never updated and never fixed. Consequence: if the stale/wrong docs live OUTSIDE the markers, no amount of reinstalling helps — verify marker position with grep before assuming a reinstall will fix drift. Also note templates are usually string literals baked into the binary rather than fetched from the server, so package upgrade alone changes nothing; the npm postinstall often just prints a nudge, and you must re-run the install subcommand after upgrading.